acceptodds
Withdrawn from ICLR 2027

WaypointVLA: Training-Efficient Vision-Language-Action Learning via Sparse Action Waypoints

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-language-action (VLA) policies typically predict dense action chunks over a long, fixed future horizon, requiring supervision for every intermediate action. However, under closed-loop control, only a small portion of each predicted chunk is executed before replanning, creating a prediction–execution mismatch. Our analysis of a policy on LIBERO further shows that distant future actions become increasingly unreliable, questioning the necessity of dense long-horizon supervision. Meanwhile, within the practically useful horizon, action trajectories can be effectively reconstructed from only a sparse set of key actions using simple linear interpolation. Motivated by these, we introduce WaypointVLA, which predicts only two key actions—a start-point action and an endpoint action—while lightweight interpolation reconstructs the intermediate low-level controls. To support training, we derive sparse waypoint supervision directly from existing demonstrations and use it to fine-tune the original VLA model without modifying the base VLA backbone. By concentrating supervision on temporally informative actions rather than redundant intermediate controls, WaypointVLA achieves faster convergence and improved robustness to visual and environmental perturbations. We evaluate WaypointVLA on LIBERO and LIBERO-Plus using as the base model. For clarity, we refer to the original baseline as the dense policy. Remarkably, using only 200 demonstrations (13.9% of the full-data setting), WaypointVLA surpasses the full-data dense policy on both benchmarks while achieving a training speedup. These results suggest that sparse waypoint training provides a more efficient alternative to dense training.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Robot Learning, Efficient Training

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ewTeOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

WaypointVLA focuses on the efficiency of VLA models. It predicts sparse waypoints and then uses interpolation to produce an action sequence, achieving a large speedup. However, this approach may reduce generalization and robustness, and it does not appear to scale well.

Critical strengths
  1. Efficiency is a key issue for VLA models. The paper explores a design for the action expert that generates trajectories through sparse prediction followed by interpolation.

  2. The paper includes a wide range of robustness tests under different environmental conditions.

  3. The writing and figures are easy to follow.

Critical weaknesses
  1. The sparse prediction approach does not show that longer-horizon prediction is unnecessary. Predicting over longer horizons is essential for more complex tasks, but the paper does not evaluate its method on such tasks.

  2. Interpolation can work well for simple tasks in relatively static environments. However, the method may struggle when the environment changes or when unexpected events during execution require corrections. The robustness evaluation should focus on changes during task execution, rather than only testing different static environment settings.

  3. The Duration Head is trained on the full dataset of 1.4k samples, while the paper claims to use only 200 demonstrations. To support this claim, all components should be trained using only those 200 demonstrations, and the comparisons should use the same data budget.

Questions for authors

Please refer to the weaknesses section. I am willing to raise the score if the authors can address my concerns.

4 comments

W1. Long-horizon prediction.

A1: We thank the reviewer for this valuable suggestion. We would like to clarify that our method does not claim that longer-horizon prediction is unnecessary, and we have clarified this point in the revised abstract. Instead, our primary motivation is to reduce redundant training costs caused by the prediction–execution mismatch in dense action prediction. Our key observation is that intermediate actions within relatively smooth segments can be reconstructed through linear interpolation. Moreover, WaypointVLA employs an adaptive Duration Head, allowing the execution horizon to vary rather than being restricted to a fixed short duration.

1. Adaptive Duration Analysis. To examine this adaptive behavior, we analyzed the predicted duration distribution during LIBERO evaluation. The results are summarized below.

SuiteMean ± StdMedianP95Range
Spatial2.731 ± 0.880341–7
Object2.912 ± 1.049351–7
Goal2.884 ± 1.088351–9
Long3.906 ± 1.437462–9
Overall3.252 ± 1.301361–9

We observe that LIBERO-Long exhibits a larger average predicted duration (3.91 steps) than the other three suites (2.73–2.91 steps), with a 95th percentile of 6 steps. This indicates that WaypointVLA can adaptively predict longer execution segments for long-horizon tasks rather than always relying on short, fixed-duration predictions.

2. Long-horizon Task Evaluation. To further address the reviewer's concern, we conducted a task-level comparison between WaypointVLA and the original π0.5\pi_{0.5} dense policy on all LIBERO-Long tasks.

TaskWaypointVLA (%)π0.5\pi_{0.5} Dense (%)
0. Alphabet soup and tomato sauce into basket8060
1. Cream cheese and butter into basket100100
2. Open stove and place moka pot10070
3. Place black bowl in drawer and close it8090
4. Place two cups on separate plates10080
5. Place book in rear compartment9070
6. Place cup and pudding at specified locations9080
7. Alphabet soup and cream cheese into basket100100
8. Place two moka pots on stove6070
9. Place cup in microwave and close door9070
Overall89.079.0

WaypointVLA achieves 89.0% overall success, compared with 79.0% for the original π0.5\pi_{0.5} dense policy, showing a 10.0-percentage-point improvement. In particular, WaypointVLA achieves higher success rates on 6 of the 10 tasks, comparable performance on 2 tasks, and lower performance on the remaining 2 tasks.

These results suggest that sparse waypoint prediction can remain effective on multi-stage, long-horizon manipulation tasks. Rather than eliminating long-horizon reasoning or control, our method reduces redundant dense action supervision while allowing the execution duration to adapt to different segments.

W2. Robustness to dynamic perturbations during execution.

A2: We thank the reviewer for this valuable suggestion. In our original experiments, we evaluated WaypointVLA and the π0.5\pi_{0.5} dense policy on LIBERO-Plus under seven types of perturbations. As shown below, WaypointVLA consistently outperforms the dense baseline across all demonstration budgets, demonstrating improved robustness to visual and environmental changes.

MethodEpisodesCameraRobotLanguageLightBackgroundNoiseLayoutTotal
Dense20053.163.571.380.080.468.073.368.9
WaypointVLA20065.270.376.889.383.379.278.676.7
Dense40055.766.771.781.078.971.473.870.5
WaypointVLA40068.872.177.985.185.777.980.777.6
Dense60058.266.673.984.381.474.273.972.3
WaypointVLA60069.074.979.489.489.180.082.779.8
Dense1,43956.367.677.383.483.674.875.573.2
WaypointVLA1,43973.375.382.895.393.785.484.083.3

To address the reviewer's concern that these perturbations are primarily static, we further conducted experiments with perturbations introduced during task execution, in both real-world and simulated environments.

1. Real-world dynamic lighting perturbations. In our real-world experiments, we randomly changed the environmental illumination using a handheld flashlight while the robot was executing the task. Each model was independently evaluated over 100 trials. Under these dynamic lighting changes, WaypointVLA achieved 82% success, compared with 74% for the π0.5\pi_{0.5} dense policy, showing an 8-percentage-point improvement.

2. Dynamic perturbations in LIBERO. We additionally evaluated both models on the 10 LIBERO-Goal tasks with Light, Background, and Noise perturbations introduced during execution. Specifically, perturbations were randomly applied at selected time points (0.3, 1.0, 1.7, and 2.4 seconds), each lasting 0.5 seconds. We varied the number, combinations, and temporal order of perturbations to simulate changing visual conditions during task execution. To ensure a fair comparison, both models were evaluated under matched perturbation distributions, with 100 trials per perturbation category for each model.

MethodLightBackgroundNoiseOverall
π0.5\pi_{0.5} Dense FT91.094.092.092.33
WaypointVLA95.094.094.094.33

WaypointVLA achieves 94.33% overall success, compared with 92.33% for the dense baseline, improving performance by 2.0 percentage points. It achieves higher success rates under dynamic lighting and noise perturbations, while maintaining comparable performance under background changes.

These additional results suggest that WaypointVLA remains effective when visual perturbations occur during task execution, rather than only under static environmental changes. While the current experiments focus on visual disturbances, they provide further evidence that sparse waypoint prediction and interpolation can maintain robust performance under dynamically changing observations.

W3. Data-budget mismatch in Duration Head training

A3: We thank the reviewer for identifying this issue. We acknowledge that sharing a Duration Head trained on the full 1,439 demonstrations across different low-data settings was an oversight in our experimental protocol.

To address this concern, we retrained separate Duration Heads for the 200-, 400-, and 600-demonstration settings, ensuring that each Duration Head only uses samples derived from the corresponding training subset. We have also updated the corresponding results, tables, and figures in the revised manuscript accordingly.

We then reevaluated the updated models on both LIBERO and LIBERO-Plus.

1. Updated LIBERO-Plus results.

EpisodesMethodCameraRobotLanguageLightBackgroundNoiseLayoutTotal
200Dense53.163.571.380.080.468.073.368.9
Waypt.65.270.376.889.383.379.278.676.7
400Dense55.766.771.781.078.971.473.870.5
Waypt.68.872.177.985.185.777.980.777.6
600Dense58.266.673.984.381.474.273.972.3
Waypt.69.074.979.489.489.180.082.779.8
1,439Dense56.367.677.383.483.674.875.573.2
Waypt.73.375.382.895.393.785.484.083.3

Compared with the original results, the budget-matched Duration Heads lead to some performance decreases, particularly in the low-data setting:

EpisodesOriginal Total (%)Updated Total (%)Δ (pp)
20078.9076.72−2.18
40078.9077.63−1.27
60080.0079.83−0.17

Nevertheless, WaypointVLA still consistently outperforms the corresponding dense baseline across all demonstration budgets, achieving improvements of +7.8, +7.1, and +7.5 percentage points with 200, 400, and 600 demonstrations, respectively.

2. Updated LIBERO results.

We also reevaluated the models on LIBERO using the budget-matched Duration Heads:

EpisodesOriginal Best SR (%)Updated Best SR (%)Δ (pp)
20091.5090.25−1.25
40094.0093.50−0.50
60094.0094.50+0.50

We observe that the performance difference generally decreases as the training budget increases, suggesting that Duration Head estimation benefits from additional demonstrations.

More importantly, even when the Duration Head is trained using only 200 demonstrations, WaypointVLA achieves 90.25% on LIBERO and 76.72% on LIBERO-Plus, outperforming the corresponding dense baselines (88.25% and 68.9%, respectively).

These additional experiments confirm that the data-efficiency advantage of WaypointVLA remains under strictly matched demonstration budgets, without requiring the Duration Head to access additional demonstrations.

AuthorsOpenReview ↗

Polite follow-up: request for reviewers’ feedback on our responses!

Hi, we sincerely thank you very much for these constructive comments and evaluation of our manuscript. We would like to kindly ask you to take a look at our responses and reevaluate our work based on our clarifications. Please let us know whether our response addresses your concerns or whether there is any further detail we can provide to help address these concerns.

Thank you again for dedicating your time to reviewing our paper.

Reviewer_YU14OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies whether a VLA policy needs dense supervision over a long action horizon when only a short prefix is executed before replanning. WaypointVLA replaces the dense action target with two action commands defining a segment, reconstructs intermediate arm commands by linear interpolation, and switches the gripper command at the segment midpoint. Training targets are obtained from existing demonstrations using action-based anchor selection followed by temporal supplementation. The policy learns from both segment boundaries and intermediate demonstration states. A separately trained duration classifier determines how long each predicted segment is executed.

Using pi_0.5, the paper reports improvements from 90.3% to 94.8% on LIBERO and from 73.2% to 83.3% on LIBERO-Plus in the full-data setting, together with policy-training speedups of approximately 2.4–2.6 times at matched demonstration budgets. Experiments with SmolVLA and an ALOHA pick-and-place task provide additional evidence. The results are promising, although the duration classifier's access to the full demonstration set complicates the low-data claims.

Critical strengths

The method is simple and addresses a practical question: how much dense action supervision does a VLA actually need? Predicting two action waypoints keeps the approach easy to integrate with existing models, and the training savings make it potentially useful.

The robustness results are encouraging, particularly on LIBERO-Plus. The additional backbone and real-world experiments give some confidence that the benefit is not specific to a single model or simulation setting.

Critical weaknesses

1: The low-data claim needs clarification. The duration predictor appears to be trained on the full demonstration set even when the main policy uses a smaller subset. If so, the complete method has access to more task data than the stated budget suggests, which weakens one of the paper's main claims.

2: The comparisons leave some uncertainty about where the gains come from. Dense-H2 does not match the waypoint policy's execution horizon, so a simpler short-horizon dense baseline is still missing. The connection to earlier waypoint extraction methods also needs a clearer discussion to establish what is new here.

Questions for authors

1: Was the duration predictor trained only on the demonstrations available in each low-data setting?

2: Does the advantage remain against a dense policy with the same execution horizon and training budget?

3: How does the proposed waypoint selection differ from AWE?

4 comments

W1 & Q1. Data-budget mismatch in Duration Head training

A1: We thank the reviewer for identifying this issue. We acknowledge that sharing a Duration Head trained on the full 1,439 demonstrations across different low-data settings was an oversight in our experimental protocol.

To address this concern, we retrained separate Duration Heads for the 200-, 400-, and 600-demonstration settings, ensuring that each Duration Head only uses samples derived from the corresponding training subset. We have also updated the corresponding results, tables, and figures in the revised manuscript accordingly.

We then reevaluated the updated models on both LIBERO and LIBERO-Plus.

1. Updated LIBERO-Plus results.

EpisodesMethodCameraRobotLanguageLightBackgroundNoiseLayoutTotal
200Dense53.163.571.380.080.468.073.368.9
Waypt.65.270.376.889.383.379.278.676.7
400Dense55.766.771.781.078.971.473.870.5
Waypt.68.872.177.985.185.777.980.777.6
600Dense58.266.673.984.381.474.273.972.3
Waypt.69.074.979.489.489.180.082.779.8
1,439Dense56.367.677.383.483.674.875.573.2
Waypt.73.375.382.895.393.785.484.083.3

Compared with the original results, the budget-matched Duration Heads lead to some performance decreases, particularly in the low-data setting:

EpisodesOriginal Total (%)Updated Total (%)Δ (pp)
20078.9076.72−2.18
40078.9077.63−1.27
60080.0079.83−0.17

Nevertheless, WaypointVLA still consistently outperforms the corresponding dense baseline across all demonstration budgets, achieving improvements of +7.8, +7.1, and +7.5 percentage points with 200, 400, and 600 demonstrations, respectively.

2. Updated LIBERO results.

We also reevaluated the models on LIBERO using the budget-matched Duration Heads:

EpisodesOriginal Best SR (%)Updated Best SR (%)Δ (pp)
20091.5090.25−1.25
40094.0093.50−0.50
60094.0094.50+0.50

We observe that the performance difference generally decreases as the training budget increases, suggesting that Duration Head estimation benefits from additional demonstrations.

More importantly, even when the Duration Head is trained using only 200 demonstrations, WaypointVLA achieves 90.25% on LIBERO and 76.72% on LIBERO-Plus, outperforming the corresponding dense baselines (88.25% and 68.9%, respectively).

These additional experiments confirm that the data-efficiency advantage of WaypointVLA remains under strictly matched demonstration budgets, without requiring the Duration Head to access additional demonstrations.

W2 & Q2. Comparison with Dense-H10 and matched execution horizons

A2: We thank the reviewer for this suggestion. To better isolate the contribution of sparse waypoint supervision, we conducted additional experiments with Dense-H10 and Dense-H50 under the Full1439 setting, including dense training on waypoint-selected observations. The corresponding experimental results have also been incorporated into the revised manuscript.

To provide a more comparable execution setting, we further analyzed the outputs of our Duration Head and found that its predicted duration is typically around 3–4 steps. Accordingly, for Dense-H10, we additionally evaluate a setting that executes the first four actions before replanning.

The results are summarized below:

MethodSpatialObjectGoalLongOverall
Dense-H50 (Exec10)97.096.082.075.087.50
Dense-H10 (Exec4)100.095.085.085.091.25
WaypointVLA99.099.092.089.094.75

Under the shorter execution setting, WaypointVLA outperforms Dense-H10 by 3.50 percentage points and Dense-H50 by 7.25 percentage points in overall success rate.

These results provide additional evidence that sparse waypoint supervision can achieve stronger performance than dense action prediction under comparable execution settings, while substantially reducing training cost.

W3 & Q3. Comparison with Automatic Waypoint Extraction (AWE)

A3: We thank the reviewer for highlighting AWE, which shares the general idea of extracting sparse waypoints from demonstrations. However, our method differs substantially in waypoint construction, training objectives, and execution. We have also added a discussion of these methodological differences in the revised manuscript.

1. Waypoint construction. AWE uses dynamic programming to identify the minimum number of waypoints that reconstruct a trajectory within a specified error threshold. In contrast, our method aims to preserve action-critical events while avoiding excessively long execution segments. We therefore adopt a two-stage strategy: (i) action-aware selection, which preserves gripper transitions and selects additional anchors based on action reconstruction error; and (ii) temporal-coverage supplementation, which repeatedly bisects the longest segments until reaching the waypoint budget.

2. Training objectives. AWE relabels observations with future waypoints while retaining the original imitation learning framework, such as ACT or Diffusion Policy. In chunk-based implementations, the model still predicts a fixed-length sequence of waypoint targets, which may contain repeated anchors, e.g., [w1,w1,w2,w2,w2,w3,…][w_1,w_1,w_2,w_2,w_2,w_3,\ldots]. In contrast, WaypointVLA treats each waypoint segment as an independent training unit and predicts only two actions, [astart,aend][a_{\mathrm{start}},a_{\mathrm{end}}], regardless of the segment length. A lightweight Duration Head further estimates the temporal span DD, allowing intermediate actions to be reconstructed through interpolation. Thus, unlike AWE's waypoint-relabeled chunk prediction, our method directly replaces dense action supervision with sparse two-action prediction and adaptive execution duration.

3. Inference and execution. AWE relies on a low-level position controller to reach predicted waypoints. Our method instead predicts two actions directly in the original VLA action space, reconstructs the intermediate control sequence through duration-conditioned interpolation, and replans after executing the segment. Thus, our primary contribution is not waypoint extraction alone, but replacing dense long-horizon VLA supervision with sparse, segment-level action prediction.

Regarding the baseline, we agree that an adapted AWE implementation provides a meaningful comparison. Specifically, AWE's waypoint selection algorithm can replace our two-stage selector while retaining the same VLA backbone, training budget, two-action prediction objective, and execution protocol. Such a comparison isolates the effect of waypoint selection, although it constitutes an AWE-adapted baseline rather than a faithful reproduction of the original AWE policy.

Furthermore, we adapted AWE to the π0.5\pi_{0.5} backbone and evaluated it under different demonstration budgets. The results are summarized below.

EpisodesMethodSpatialObjectGoalLongOverall
200AWE + π0.5\pi_{0.5}27.066.034.01.032.00
WaypointVLA99.095.088.079.090.25
400AWE + π0.5\pi_{0.5}29.084.034.012.039.75
WaypointVLA99.099.091.085.093.50
600AWE + π0.5\pi_{0.5}51.074.034.011.042.50
WaypointVLA100.099.088.091.094.50
1,439AWE + π0.5\pi_{0.5}65.074.049.024.053.00
WaypointVLA99.099.092.089.094.75

In AWE training, the model still predicts fixed-length dense chunks, with the original action targets replaced by waypoint targets. Therefore, it does not reduce the number of supervised action positions or provide the computational savings of our sparse two-action prediction. Our results suggest that the proposed sparse two-action prediction framework is more effective and achieves better performance than the AWE-adapted approach under the evaluated settings.

AuthorsOpenReview ↗

Polite follow-up: request for reviewers’ feedback on our responses!

Hi, we sincerely thank you very much for these constructive comments and evaluation of our manuscript. We would like to kindly ask you to take a look at our responses and reevaluate our work based on our clarifications. Please let us know whether our response addresses your concerns or whether there is any further detail we can provide to help address these concerns.

Thank you again for dedicating your time to reviewing our paper.

Reviewer_bf9yOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces WaypointVLA, a training-efficient approach that replaces dense action-chunk prediction with two action waypoints and reconstructs intermediate controls through linear interpolation. The method extracts waypoint supervision from existing demonstrations using action-aware anchor selection and temporal coverage, and optionally predicts segment duration with a lightweight classifier. Experiments on LIBERO and LIBERO-Plus show improved success rates and approximately 2.4–2.6× lower training time under matched demonstration budgets. Additional experiments with SmolVLA and a real robot provide evidence of transfer beyond the primary backbone and simulation setting.

Critical strengths

Simple and practical formulation. The method derives sparse supervision from existing demonstrations and preserves the VLA backbone, making it relatively straightforward to incorporate into existing training pipelines. Consistent empirical improvements. WaypointVLA outperforms the matched dense baseline across all evaluated demonstration budgets on both benchmarks. The full-data improvement on LIBERO-Plus, from 73.2% to 83.3%, is particularly encouraging. Supporting evidence across settings. Experiments with a second backbone and real-world manipulation strengthen the empirical contribution. The duration and boundary–recovery ablations also provide useful insights into the design.

Critical weaknesses

Data-budget mismatch: The duration head uses all 1,439 demonstrations, weakening the claim that the complete system requires only 200 demonstrations. Insufficient controls: Dense-H2 does not isolate waypoint supervision from differences in sample selection, temporal coverage, and replanning frequency. Efficiency attribution: Most training savings appear to come from fewer optimizer updates. The headline 15.2× speedup also compares different demonstration budgets. Limited validation: Single-seed results and the missing comparison with Automatic Waypoint Extraction leave uncertainty about robustness and novelty.

Questions for authors

Is the full-data duration head shared across all demonstration budgets? How does performance change when it is trained only on the corresponding subset? How does WaypointVLA compare with Dense-H10 and a dense policy trained on the same selected observations under matched compute budgets? Do the reported speedups include waypoint extraction, simulator replay, and duration-head training? How are checkpoints selected, and do the improvements persist across multiple training seeds and demonstration subsets? What are the key methodological differences from Automatic Waypoint Extraction, and could an adapted version serve as a baseline?

7 comments

W1.1 & Q1–Q2. Data-budget mismatch in Duration Head training

A1: We thank the reviewer for identifying this issue. We acknowledge that sharing a Duration Head trained on the full 1,439 demonstrations across different low-data settings was an oversight in our experimental protocol.

To address this concern, we retrained separate Duration Heads for the 200-, 400-, and 600-demonstration settings, ensuring that each Duration Head only uses samples derived from the corresponding training subset. We have also updated the corresponding results, tables, and figures in the revised manuscript accordingly.

We then reevaluated the updated models on both LIBERO and LIBERO-Plus.

1. Updated LIBERO-Plus results.

EpisodesMethodCameraRobotLanguageLightBackgroundNoiseLayoutTotal
200Dense53.163.571.380.080.468.073.368.9
Waypt.65.270.376.889.383.379.278.676.7
400Dense55.766.771.781.078.971.473.870.5
Waypt.68.872.177.985.185.777.980.777.6
600Dense58.266.673.984.381.474.273.972.3
Waypt.69.074.979.489.489.180.082.779.8
1,439Dense56.367.677.383.483.674.875.573.2
Waypt.73.375.382.895.393.785.484.083.3

Compared with the original results, the budget-matched Duration Heads lead to some performance decreases, particularly in the low-data setting:

EpisodesOriginal Total (%)Updated Total (%)Δ (pp)
20078.9076.72−2.18
40078.9077.63−1.27
60080.0079.83−0.17

Nevertheless, WaypointVLA still consistently outperforms the corresponding dense baseline across all demonstration budgets, achieving improvements of +7.8, +7.1, and +7.5 percentage points with 200, 400, and 600 demonstrations, respectively.

2. Updated LIBERO results.

We also reevaluated the models on LIBERO using the budget-matched Duration Heads:

EpisodesOriginal Best SR (%)Updated Best SR (%)Δ (pp)
20091.5090.25−1.25
40094.0093.50−0.50
60094.0094.50+0.50

We observe that the performance difference generally decreases as the training budget increases, suggesting that Duration Head estimation benefits from additional demonstrations.

More importantly, even when the Duration Head is trained using only 200 demonstrations, WaypointVLA achieves 90.25% on LIBERO and 76.72% on LIBERO-Plus, outperforming the corresponding dense baselines (88.25% and 68.9%, respectively).

These additional experiments confirm that the data-efficiency advantage of WaypointVLA remains under strictly matched demonstration budgets, without requiring the Duration Head to access additional demonstrations.

W1.2 & Q3. Comparison with Dense-H10 and matched execution horizons

A2: We thank the reviewer for this suggestion. To better isolate the contribution of sparse waypoint supervision, we conducted additional experiments with Dense-H10 and Dense-H50 under the Full1439 setting, including dense training on waypoint-selected observations. The corresponding experimental results have also been incorporated into the revised manuscript.

To provide a more comparable execution setting, we further analyzed the outputs of our Duration Head and found that its predicted duration is typically around 3–4 steps. Accordingly, for Dense-H10, we additionally evaluate a setting that executes the first four actions before replanning.

The results are summarized below:

MethodSpatialObjectGoalLongOverall
Dense-H50 (Exec10)97.096.082.075.087.50
Dense-H10 (Exec4)100.095.085.085.091.25
WaypointVLA99.099.092.089.094.75

Under the shorter execution setting, WaypointVLA outperforms Dense-H10 by 3.50 percentage points and Dense-H50 by 7.25 percentage points in overall success rate.

These results provide additional evidence that sparse waypoint supervision can achieve stronger performance than dense action prediction under comparable execution settings, while substantially reducing training cost.

W1.3. Efficiency attribution

A3: We thank the reviewer for this valuable comment. We agree that the training speedup primarily comes from the reduction in training samples and corresponding optimizer updates, enabled by sparse waypoint supervision.

Instead of supervising every intermediate action, WaypointVLA only supervises the segment endpoints, substantially reducing the number of training samples and optimizer updates. Meanwhile, the adaptive Duration Head allows the policy to reconstruct action sequences of different lengths, preserving task performance despite the reduced supervision density.

We also acknowledge that the reported 15.2× speedup compares different demonstration budgets. Our intention was to highlight that WaypointVLA can achieve better performance with fewer demonstrations and substantially less training time, as demonstrated on both LIBERO and LIBERO-Plus.

To avoid ambiguity, we have revised the abstract to emphasize the 2.46× training speedup under the same 200-demonstration budget, rather than the 15.2× cross-budget comparison.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.